GGUF builds of nokeepai/Inkling-2-Mini-Unbound, an uncensored Inkling 2 Mini. The full model card has the method, evaluations and limits.
Built by nokeep, who also sponsored the compute. nokeep runs private, uncensored AI chat, image and video at nokeep.ai.
This model follows harmful instructions. Its refusal behaviour has been removed. Use it responsibly and in line with the law where you are.
Measured on a sealed test split against the original Inkling-2-Mini (full evaluation in the main model card):
| Original | Unbound | |
|---|---|---|
| Harmful prompts refused (thinking off / effort 0.9) | 97.5% / 97.5% | 0% / 0% |
| Useful answers to harmful requests (off / 0.9) | 0.4% / 1.1% | 54% / 63% |
| Benign borderline prompts refused | 38% | 0% |
| MMLU-Pro / GSM8K / MATH-500 (subsets) | 53.5 / 96.6 / 96.5 | 52.5 / 96.4 / 96.0 |
| GPQA Diamond / IFEval / HumanEval+ | 68.7 / 90.6 / 90.2 | 74.2 / 89.7 / 92.1 |
These numbers come from the safetensors model on the sealed test split. The GGUF files carry the same edit; see the KLD column below for how close each size stays to it.
Share of harmful requests that get a genuinely useful answer (Jev judge, 95% CIs).
Capability benchmarks with 95% CIs against the original's range.
| File | Size | KLD vs original BF16 | Same top token |
|---|---|---|---|
| Inkling-2-Mini-Unbound-Q8_0.gguf | 27.8 GB | 0.0118 | 95.2% |
| Inkling-2-Mini-Unbound-Q6_K.gguf | 23.5 GB | 0.0127 | 95.1% |
| Inkling-2-Mini-Unbound-Q5_K_M.gguf | 20.0 GB | 0.0139 | 94.8% |
| Inkling-2-Mini-Unbound-Q4_K_M.gguf | 17.6 GB | 0.0188 | 94.0% |
| Inkling-2-Mini-Unbound-IQ3_M.gguf | 13.0 GB | 0.0402 | 91.3% |
| Inkling-2-Mini-Unbound-IQ2_M.gguf | 10.8 GB | 0.0828 | 88.0% |
| Inkling-2-Mini-Unbound-BF16.gguf | 52.2 GB | 0.0116 | 95.2% |
| mmproj-Inkling-2-Mini-BF16.gguf / -Q8_0.gguf | 0.6 / 0.3 GB | vision projector, unchanged from the original | |
| Inkling-2-Mini-Unbound-LoRA-r16.gguf | 7 MB | LoRA adapter for any GGUF of the original (see below) |
For reference, the original model's own Q8_0 measures 0.0036 and its IQ4_NL 0.0096 on the same test, so about 0.012 of the numbers above is the uncensoring edit itself and the rest is quantization. KLD and top-token agreement are measured against the original BF16 GGUF on 40 chunks of wikitext-2 (512 tokens each).
Q4_K_M is the size most people want. Q8_0 and Q6_K are close to lossless. IQ3_M and IQ2_M are for small GPUs and are noticeably lossy.
These start from the official ggml-org/Inkling-2-Mini-GGUF BF16 file. Only our 20 edited attention-output matrices (blk.4-23.attn_output.weight) were swapped in; the other 684 tensors are byte-identical to ggml-org's file (per-tensor hashes in MANIFEST.tsv). Every smaller size was then quantized with ggml-org's importance matrix, keeping those 20 edited matrices at Q8_0 so the edit isn't blurred by 4-bit compression. Every file refused 0 of 10 harmful test prompts, where the original refuses 9 of 10.
You need a llama.cpp build with Inkling 2 loop support that includes the loop fix from PR #32704. That fix is on master (we used commit 19bc41dbb); tagged releases up to b12608 don't have it yet. Apps that bundle llama.cpp (LM Studio, Ollama, etc.) need an equally recent build.
./build/bin/llama-server \
-m Inkling-2-Mini-Unbound-Q4_K_M.gguf \
--mmproj mmproj-Inkling-2-Mini-Q8_0.gguf \
--jinja -c 32768 -ngl 99
If you already have a GGUF of the original Inkling 2 Mini (for example ggml-org's), load the adapter instead of downloading a whole new model:
./build/bin/llama-server \
-m Inkling-2-Mini-Q8_0.gguf \
--lora Inkling-2-Mini-Unbound-LoRA-r16.gguf \
--mmproj mmproj-Inkling-2-Mini-Q8_0.gguf \
--jinja -c 32768 -ngl 99
The LoRA is a rank-16 approximation of the edit, not a bit-exact copy: our edit also rescales each row, which a plain LoRA can't express exactly. In practice it behaves like the full model. On ggml-org's Q8_0 it refused 0 of 10 harmful test prompts (the original refuses 9), and its KLD against our full Q8_0 is 0.0029, smaller than the original's own Q8_0-vs-BF16 difference (0.0036). The base GGUF must use the merged Inkling 2 loop layout (llama.cpp PR #32694).
reasoning_effort, 0 to 1), the same as the original. The model still thinks briefly even at 0.-ub 16, or load the whole model on GPU.This model was built, and its compute paid for, by nokeep. nokeep is a private AI app for uncensored chat, images and video. If this model is useful to you, the best way to support more releases like it is to try nokeep.